Optimization and assessment of Omega-3 fatty acids from high-dimensional spectroscopic data in the Atlantic salmon breeding programs

Atlantic salmon is well known as a rich source of Omega-3 fatty acids (in particular, ALA, DHA, and EPA), and these fatty acids are heritable, thus, the selection of parents will influence their levels in the offspring. The gold standard method of recording Omega-3 fatty acids in Atlantic salmon is costly, time-consuming, and destructive to the sample. For selective breeding purposes, a more affordable method is needed to measure Omega-3 fatty acids in thousands of related salmon. In many breeding programs, vibrational spectroscopy is primarily used with the Partial Least Squares Regression (PLSR) model to measure and predict phenotypes such as Omega-3 fatty acids. However, there is a knowledge gap in estimating heritability using vibrational spectroscopy and finding the effect of sample selection and variable selection methods in the data analysis process perspective. Hence, we optimized and assessed the predicted Omega-3 fatty acids from the high-dimensional spectroscopy data (NIR and Raman spectroscopy data) in the breeding program according to the multiple scenarios combined with sample selection (Kennard-Stone and Random Sampling) and variable selection (with or without Markov Blanket). We found that the PLSR model accuracy and the resulting heritability estimates generally increase with adopting the variable selection method. We also found that NIR spectroscopy has a good affinity with Kennard-Stone sampling, while Raman spectroscopy showed stable performance regardless of sample selection. It is possible to achieve improved PLSR model accuracy and heritability estimates by utilizing the Markov Blanket approach.